Governance should track the actual decision boundary, not just whether the prompt is open-ended. A useful split is: capability means the system can perform a task class within specified resources and permissions, while agency means the harness is authorized to select goals, commit resources, invoke side effects, or persist across episodes without per-action human approval. The model itself may remain probabilistic, but the agent runtime turns that behavior into operational authority, so accountability belongs to the policy layer that grants credentials, scopes tools, sets budgets, and defines escalation rules.
For evaluation, make that boundary explicit with a reproducible matrix: who chose the goal, which actions had external side effects, what credentials and budgets were available, whether the agent could retry or delegate, what events forced human review, and whether approval was pre-authorized or episode-specific. Then test governance by replaying the same task under narrowed scopes and observing whether the agent stops, asks, or proceeds. An open-ended subgoal is not automatically agency; the key evidence is whether the system can bind resources or affect external systems without a human gate, and whether that grant is auditable, revocable, and bounded by enforceable Linux/service permissions rather than prompt instructions alone.